OLD | NEW |
---|---|
1 /* | 1 /* |
2 * Copyright 2016 Google Inc. | 2 * Copyright 2016 Google Inc. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license that can be | 4 * Use of this source code is governed by a BSD-style license that can be |
5 * found in the LICENSE file. | 5 * found in the LICENSE file. |
6 */ | 6 */ |
7 | 7 |
8 #ifndef Sk4x4f_DEFINED | 8 #ifndef Sk4x4f_DEFINED |
9 #define Sk4x4f_DEFINED | 9 #define Sk4x4f_DEFINED |
10 | 10 |
11 #include "SkNx.h" | 11 #include "SkNx.h" |
12 | 12 |
13 struct Sk4x4f { | 13 struct Sk4x4f { |
14 Sk4f r,g,b,a; | 14 Sk4f r,g,b,a; |
15 | 15 |
16 static Sk4x4f Transpose(const Sk4f&, const Sk4f&, const Sk4f&, const Sk4f&); | 16 static Sk4x4f Transpose(const Sk4f&, const Sk4f&, const Sk4f&, const Sk4f&); |
17 static Sk4x4f Transpose(const float[16]); | 17 static Sk4x4f Transpose(const float[16]); |
18 static Sk4x4f Transpose(const uint8_t[16]); | 18 static Sk4x4f Transpose(const uint8_t[16]); |
19 | 19 |
20 void transpose(Sk4f*, Sk4f*, Sk4f*, Sk4f*) const; | 20 void transpose(Sk4f* x, Sk4f* y, Sk4f* z, Sk4f* w) const { |
21 auto t = Transpose(r,g,b,a); | |
22 *x = t.r; | |
23 *y = t.g; | |
24 *z = t.b; | |
25 *w = t.a; | |
26 } | |
21 void transpose( float[16]) const; | 27 void transpose( float[16]) const; |
22 void transpose(uint8_t[16]) const; | 28 void transpose(uint8_t[16]) const; |
23 }; | 29 }; |
24 | 30 |
25 // TODO: NEON | |
26 | |
27 #if 1 && !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 | 31 #if 1 && !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2 |
28 | 32 |
29 inline Sk4x4f Sk4x4f::Transpose(const Sk4f& x, const Sk4f& y, const Sk4f& z, con st Sk4f& w) { | 33 inline Sk4x4f Sk4x4f::Transpose(const Sk4f& x, const Sk4f& y, const Sk4f& z, con st Sk4f& w) { |
30 auto r = x.fVec, | 34 auto r = x.fVec, |
31 g = y.fVec, | 35 g = y.fVec, |
32 b = z.fVec, | 36 b = z.fVec, |
33 a = w.fVec; | 37 a = w.fVec; |
34 _MM_TRANSPOSE4_PS(r,g,b,a); | 38 _MM_TRANSPOSE4_PS(r,g,b,a); |
35 return { r,g,b,a }; | 39 return { r,g,b,a }; |
36 } | 40 } |
37 | 41 |
38 inline Sk4x4f Sk4x4f::Transpose(const float fs[16]) { | 42 inline Sk4x4f Sk4x4f::Transpose(const float fs[16]) { |
39 return Transpose(Sk4f::Load(fs+0), Sk4f::Load(fs+4), Sk4f::Load(fs+8), Sk4f: :Load(fs+12)); | 43 return Transpose(Sk4f::Load(fs+0), Sk4f::Load(fs+4), Sk4f::Load(fs+8), Sk4f: :Load(fs+12)); |
40 } | 44 } |
41 | 45 |
42 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) { | 46 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) { |
43 auto b16 = _mm_loadu_si128((const __m128i*)bs); | 47 auto b16 = _mm_loadu_si128((const __m128i*)bs); |
44 | 48 |
45 auto mask = _mm_set1_epi32(0xFF); | 49 auto mask = _mm_set1_epi32(0xFF); |
46 auto r = _mm_cvtepi32_ps(_mm_and_si128(mask, (b16 ))), | 50 auto r = _mm_cvtepi32_ps(_mm_and_si128(mask, (b16 ))), |
47 g = _mm_cvtepi32_ps(_mm_and_si128(mask, _mm_srli_epi32(b16, 8))), | 51 g = _mm_cvtepi32_ps(_mm_and_si128(mask, _mm_srli_epi32(b16, 8))), |
48 b = _mm_cvtepi32_ps(_mm_and_si128(mask, _mm_srli_epi32(b16, 16))), | 52 b = _mm_cvtepi32_ps(_mm_and_si128(mask, _mm_srli_epi32(b16, 16))), |
49 a = _mm_cvtepi32_ps( _mm_srli_epi32(b16, 24)); | 53 a = _mm_cvtepi32_ps( _mm_srli_epi32(b16, 24)); |
50 return { r,g,b,a }; | 54 return { r,g,b,a }; |
51 } | 55 } |
52 | 56 |
53 inline void Sk4x4f::transpose(Sk4f* x, Sk4f* y, Sk4f* z, Sk4f* w) const { | |
54 auto R = r.fVec, | |
55 G = g.fVec, | |
56 B = b.fVec, | |
57 A = a.fVec; | |
58 _MM_TRANSPOSE4_PS(R,G,B,A); | |
59 *x = R; | |
60 *y = G; | |
61 *z = B; | |
62 *w = A; | |
63 } | |
64 | |
65 inline void Sk4x4f::transpose(float fs[16]) const { | 57 inline void Sk4x4f::transpose(float fs[16]) const { |
66 Sk4f x,y,z,w; | 58 Sk4f x,y,z,w; |
67 this->transpose(&x,&y,&z,&w); | 59 this->transpose(&x,&y,&z,&w); |
68 x.store(fs+ 0); | 60 x.store(fs+ 0); |
69 y.store(fs+ 4); | 61 y.store(fs+ 4); |
70 z.store(fs+ 8); | 62 z.store(fs+ 8); |
71 w.store(fs+12); | 63 w.store(fs+12); |
72 } | 64 } |
73 | 65 |
74 inline void Sk4x4f::transpose(uint8_t bs[16]) const { | 66 inline void Sk4x4f::transpose(uint8_t bs[16]) const { |
75 auto R = _mm_cvttps_epi32(r.fVec), | 67 auto R = _mm_cvttps_epi32(r.fVec), |
76 G = _mm_slli_epi32(_mm_cvttps_epi32(g.fVec), 8), | 68 G = _mm_slli_epi32(_mm_cvttps_epi32(g.fVec), 8), |
77 B = _mm_slli_epi32(_mm_cvttps_epi32(b.fVec), 16), | 69 B = _mm_slli_epi32(_mm_cvttps_epi32(b.fVec), 16), |
78 A = _mm_slli_epi32(_mm_cvttps_epi32(a.fVec), 24); | 70 A = _mm_slli_epi32(_mm_cvttps_epi32(a.fVec), 24); |
79 _mm_storeu_si128((__m128i*)bs, _mm_or_si128(A, _mm_or_si128(B, _mm_or_si128( G, R)))); | 71 _mm_storeu_si128((__m128i*)bs, _mm_or_si128(A, _mm_or_si128(B, _mm_or_si128( G, R)))); |
80 } | 72 } |
81 | 73 |
74 #elif defined(SK_ARM_HAS_NEON) | |
75 | |
76 inline Sk4x4f Sk4x4f::Transpose(const Sk4f& x, const Sk4f& y, const Sk4f& z, con st Sk4f& w) { | |
77 float32x4x2_t xy = vuzpq_f32(x.fVec, y.fVec), | |
78 zw = vuzpq_f32(z.fVec, w.fVec), | |
79 rb = vuzpq_f32(xy.val[0], zw.val[0]), | |
80 ga = vuzpq_f32(xy.val[1], zw.val[1]); | |
81 return { rb.val[0], ga.val[0], rb.val[1], ga.val[1] }; | |
82 } | |
83 | |
84 inline Sk4x4f Sk4x4f::Transpose(const float fs[16]) { | |
85 float32x4x4_t v = vld4q_f32(fs); | |
86 return { v.val[0], v.val[1], v.val[2], v.val[3] }; | |
87 } | |
88 | |
89 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) { | |
90 auto b16 = vreinterpretq_u32_u8(vld1q_u8(bs)); | |
91 auto r = vcvtq_f32_u32(vandq_u32(vdupq_n_u32(0x000000FF), b16) ), | |
msarett
2016/03/24 17:00:50
Woohoo this is cool!
| |
92 g = vcvtq_n_f32_u32(vandq_u32(vdupq_n_u32(0x0000FF00), b16), 8), | |
93 b = vcvtq_n_f32_u32(vandq_u32(vdupq_n_u32(0x00FF0000), b16), 16), | |
94 a = vcvtq_n_f32_u32(vandq_u32(vdupq_n_u32(0xFF000000), b16), 24); | |
95 return { r,g,b,a }; | |
96 } | |
97 | |
98 inline void Sk4x4f::transpose(float fs[16]) const { | |
99 float32x4x4_t v = {{ r.fVec, g.fVec, b.fVec, a.fVec }}; | |
100 vst4q_f32(fs, v); | |
101 } | |
102 | |
103 inline void Sk4x4f::transpose(uint8_t bs[16]) const { | |
104 auto R = vandq_u32(vdupq_n_u32(0x000000FF), vcvtq_u32_f32(r.fVec )), | |
105 G = vandq_u32(vdupq_n_u32(0x0000FF00), vcvtq_n_u32_f32(g.fVec, 8)), | |
106 B = vandq_u32(vdupq_n_u32(0x00FF0000), vcvtq_n_u32_f32(b.fVec, 16)), | |
107 A = vandq_u32(vdupq_n_u32(0xFF000000), vcvtq_n_u32_f32(a.fVec, 24)); | |
108 vst1q_u8(bs, vreinterpretq_u8_u32(vorrq_u32(A, vorrq_u32(B, vorrq_u32(G, R)) ))); | |
109 } | |
110 | |
82 #else | 111 #else |
83 | 112 |
84 inline Sk4x4f Sk4x4f::Transpose(const Sk4f& x, const Sk4f& y, const Sk4f& z, con st Sk4f& w) { | 113 inline Sk4x4f Sk4x4f::Transpose(const Sk4f& x, const Sk4f& y, const Sk4f& z, con st Sk4f& w) { |
85 return { | 114 return { |
86 { x[0], y[0], z[0], w[0] }, | 115 { x[0], y[0], z[0], w[0] }, |
87 { x[1], y[1], z[1], w[1] }, | 116 { x[1], y[1], z[1], w[1] }, |
88 { x[2], y[2], z[2], w[2] }, | 117 { x[2], y[2], z[2], w[2] }, |
89 { x[3], y[3], z[3], w[3] }, | 118 { x[3], y[3], z[3], w[3] }, |
90 }; | 119 }; |
91 } | 120 } |
92 | 121 |
93 inline Sk4x4f Sk4x4f::Transpose(const float fs[16]) { | 122 inline Sk4x4f Sk4x4f::Transpose(const float fs[16]) { |
94 return Transpose(Sk4f::Load(fs+0), Sk4f::Load(fs+4), Sk4f::Load(fs+8), Sk4f: :Load(fs+12)); | 123 return Transpose(Sk4f::Load(fs+0), Sk4f::Load(fs+4), Sk4f::Load(fs+8), Sk4f: :Load(fs+12)); |
95 } | 124 } |
96 | 125 |
97 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) { | 126 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) { |
98 return { | 127 return { |
99 { (float)bs[0], (float)bs[4], (float)bs[ 8], (float)bs[12] }, | 128 { (float)bs[0], (float)bs[4], (float)bs[ 8], (float)bs[12] }, |
100 { (float)bs[1], (float)bs[5], (float)bs[ 9], (float)bs[13] }, | 129 { (float)bs[1], (float)bs[5], (float)bs[ 9], (float)bs[13] }, |
101 { (float)bs[2], (float)bs[6], (float)bs[10], (float)bs[14] }, | 130 { (float)bs[2], (float)bs[6], (float)bs[10], (float)bs[14] }, |
102 { (float)bs[3], (float)bs[7], (float)bs[11], (float)bs[15] }, | 131 { (float)bs[3], (float)bs[7], (float)bs[11], (float)bs[15] }, |
103 }; | 132 }; |
104 } | 133 } |
105 | 134 |
106 inline void Sk4x4f::transpose(Sk4f* x, Sk4f* y, Sk4f* z, Sk4f* w) const { | |
107 *x = { r[0], g[0], b[0], a[0] }; | |
108 *y = { r[1], g[1], b[1], a[1] }; | |
109 *z = { r[2], g[2], b[2], a[2] }; | |
110 *w = { r[3], g[3], b[3], a[3] }; | |
111 } | |
112 | |
113 inline void Sk4x4f::transpose(float fs[16]) const { | 135 inline void Sk4x4f::transpose(float fs[16]) const { |
114 Sk4f x,y,z,w; | 136 Sk4f x,y,z,w; |
115 this->transpose(&x,&y,&z,&w); | 137 this->transpose(&x,&y,&z,&w); |
116 x.store(fs+ 0); | 138 x.store(fs+ 0); |
117 y.store(fs+ 4); | 139 y.store(fs+ 4); |
118 z.store(fs+ 8); | 140 z.store(fs+ 8); |
119 w.store(fs+12); | 141 w.store(fs+12); |
120 } | 142 } |
121 | 143 |
122 inline void Sk4x4f::transpose(uint8_t bs[16]) const { | 144 inline void Sk4x4f::transpose(uint8_t bs[16]) const { |
123 bs[ 0] = (uint8_t)r[0]; bs[ 1] = (uint8_t)g[0]; bs[ 2] = (uint8_t)b[0]; bs[ 3] = (uint8_t)a[0]; | 145 bs[ 0] = (uint8_t)r[0]; bs[ 1] = (uint8_t)g[0]; bs[ 2] = (uint8_t)b[0]; bs[ 3] = (uint8_t)a[0]; |
124 bs[ 4] = (uint8_t)r[1]; bs[ 5] = (uint8_t)g[1]; bs[ 6] = (uint8_t)b[1]; bs[ 7] = (uint8_t)a[1]; | 146 bs[ 4] = (uint8_t)r[1]; bs[ 5] = (uint8_t)g[1]; bs[ 6] = (uint8_t)b[1]; bs[ 7] = (uint8_t)a[1]; |
125 bs[ 8] = (uint8_t)r[2]; bs[ 9] = (uint8_t)g[2]; bs[10] = (uint8_t)b[2]; bs[1 1] = (uint8_t)a[2]; | 147 bs[ 8] = (uint8_t)r[2]; bs[ 9] = (uint8_t)g[2]; bs[10] = (uint8_t)b[2]; bs[1 1] = (uint8_t)a[2]; |
126 bs[12] = (uint8_t)r[3]; bs[13] = (uint8_t)g[3]; bs[14] = (uint8_t)b[3]; bs[1 5] = (uint8_t)a[3]; | 148 bs[12] = (uint8_t)r[3]; bs[13] = (uint8_t)g[3]; bs[14] = (uint8_t)b[3]; bs[1 5] = (uint8_t)a[3]; |
127 } | 149 } |
128 | 150 |
129 #endif | 151 #endif |
130 | 152 |
131 #endif//Sk4x4f_DEFINED | 153 #endif//Sk4x4f_DEFINED |
OLD | NEW |